Deep Convolutional Neural Network for Multiple Smart Grid Event Classification through Sliding Windows

The power grid is the most expansive infrastructure man has ever made. As more smart devices capable of recording Power Quality (PQ) and other events are integrated, a “deluge of data” is overwhelming power utility resources and personnel. Power utility personnel are overwhelmed by their reliance on “by-hand” analysis of PQ events. The amount of data and the time-consuming nature associated with by-hand analysis results in most PQ events being analyzed days to weeks after they occur, if at all. This limits the value of the analysis in terms of actionable information and creates a blindspot preventing timely corrective action and loss of situational awareness. Machine learning and, specifically, Deep Learning (DL) have shown to be viable approaches for learning from large data sets, thus providing a viable approach for analyzing PQ events as they occur and even predicting events before they cause an outage, damage equipment, or endanger utility personnel and bystanders. This work presents an automated PQ Disturbance (PQD) event classification approach built on sliding window data preparation and a Convolutional Neural Network (CNN). Sliding window data preparation enables the classification of PQDs associated with voltage and current waveforms of different lengths and sampling rates. Twenty-one real smart grid events, collected at three sampling rates by smart devices deployed in an operational power transmission system, are classified using a four-layer CNN at an average percent correct classification performance of 98.45%.